Closed cxxgtxy closed 4 years ago
@cxxgtxy I didn't train the s, m, l from scratch, they are fine tuned from TF weights to work with the defaultl BN eps and no SAME padding. The XL I have trained from scratch twice, once with normal color aug, the most recent time with RandAugment. Hparams for the last run are in the README, I still think it can be improved though, if I try one of those again I'll maybe slow the LR decay. I think hparams that have been reported here for training efficientnets should also be reasonable for MixNet.
Thank you.
Thanks for your impressive work! Can you share your hyper-parameters for mixnet-m (77.256% top 1 reported here.)? Google said they used the same tricks as mnas-net, I cannot reproduce that high top 1 acc following their settings.